Derive failover time from health-check interval, failure threshold, DNS TTL and draining, and price the DNS queries.
A global load balancer cannot react faster than interval × threshold, because it deliberately waits for consecutive failures to avoid ejecting an endpoint over one dropped packet. After the decision, clients keep using the old answer until the DNS TTL expires and existing connections finish draining, so the DNS TTL is usually the largest single term. Tightening the health check without lowering the DNS TTL barely improves real failover time while multiplying probe traffic against every endpoint. Query pricing and per-endpoint monitoring fees vary by provider, so confirm the current rates before you size the pool.
Global Load Balancer
detection = health-check interval × failure threshold; convergence = decision propagation + DNS TTL + connection draining; total failover = detection + convergence.
detection = health-check interval × failure threshold; convergence = decision propagation + DNS TTL + connection draining; total failover = detection + convergence. A global load balancer cannot react faster than interval × threshold, because it deliberately waits for consecutive failures to avoid ejecting an endpoint over one dropped packet. After the decision, clients keep using the old answer until the DNS TTL expires and existing connections finish draining, so the DNS TTL is usually the largest single term.
Tightening the health check without lowering the DNS TTL barely improves real failover time while multiplying probe traffic against every endpoint. Query pricing and per-endpoint monitoring fees vary by provider, so confirm the current rates before you size the pool.
This calculator takes 11 inputs: Health-check interval, Consecutive failures before ejection, Decision propagation to the edge, DNS TTL on the load-balanced record, Connection draining, Failover events per month, Endpoints in the pool, Locations running the health check, Authoritative DNS queries, Price per million DNS queries, Monthly fee per monitored endpoint. The pre-filled defaults are a realistic starting point — replace them with figures from your own environment for a result you can act on.
Because a single lost probe would then eject a healthy endpoint, and with several probe locations against every endpoint the check traffic itself becomes significant. Consecutive-failure thresholds exist to trade a few seconds of detection time for far fewer false ejections.